DRL-based Indoor Autonomous Exploration

DRL-based Indoor Autonomous Exploration

← About
📝 Detailed writeup coming soon — key information is shown below.

DRL-based Indoor Autonomous Exploration Research

Period: Dec 2025 – Present  ·  NUS CORE Lab

Tech stack: ROS2 Humble · PyTorch · Ray · OctoMap · OccupancyGrid · ARiADNE · TARE · Livox/Hesai SLAM · Unity · Gazebo

Overview

Traditional frontier-based planners like TARE struggle in structured indoor environments — narrow corridors, T-junctions, and dead-ends cause goal=NaN failures and premature exploration termination. This project replaces TARE with ARiADNE, a DRL-based autonomous explorer, integrated into an existing ROS2 navigation stack via a minimal-substitution strategy.

Failure Analysis

  • Diagnosed TARE's parameter coupling issues, goal=NaN failures, and frontier utility collapse in narrow corridor and junction scenarios
  • Mapped out the 3D point cloud SLAM → 2D OccupancyGrid perception-planning interface to identify integration points

System Integration

  • Integrated ARiADNE into the existing ROS2 navigation chain with minimal changes to surrounding components
  • Connected Livox/Hesai SLAM output, OctoMap/OccupancyGrid, and waypoint interfaces
  • Preserved the downstream local planner / path follower to minimize system disruption

Validation

  • Built smoke-test and experiment orchestration workflows in Unity and Gazebo environments
  • Pre-trained ARiADNE checkpoint completed exploration in all 5 test episodes
  • Confirmed weight compatibility between the training repository and the ROS2 inference repository

Exploration videos, quantitative comparison with TARE, and system diagrams will be added here.

📝 详细内容即将补充，以下为主要信息。

基于 DRL 的室内自主探索导航系统重构 科研

时间： 2025.12 – 至今  ·  新加坡国立大学 CORE Lab

技术栈： ROS2 Humble · PyTorch · Ray · OctoMap · OccupancyGrid · ARiADNE · TARE · Livox/Hesai SLAM · Unity · Gazebo

项目概述

传统 frontier-based 规划器（如 TARE）在结构化室内场景——窄走廊、T 型路口、死角——中容易出现 goal=NaN 失效和过早结束探索的问题。本项目以"最小替换"策略将 TARE 替换为 ARiADNE（基于 DRL 的自主探索器），将其接入现有 ROS2 导航链路。

失效分析

  • 诊断 TARE 在窄走廊和路口场景下的参数耦合问题、goal=NaN 失效与 frontier utility 崩溃问题
  • 梳理 3D 点云 SLAM → 2D 占用栅格感知-规划接口，确定集成替换点

系统集成

  • 采用"最小替换"策略将 ARiADNE 接入现有 ROS2 导航链路，对周边组件改动最小
  • 打通 Livox/Hesai SLAM 输出、OctoMap/OccupancyGrid 与 waypoint 接口
  • 保留下游 local planner / path follower，最小化系统改动

验证

  • 在 Unity 与 Gazebo 环境中建立 smoke test 与实验编排流程
  • 预训练 ARiADNE checkpoint 在 5 个测试 episode 上均完成探索任务
  • 确认训练仓与 ROS2 推理仓的权重兼容性

探索过程视频、与 TARE 的定量对比和系统结构图将在后续补充。